export

游雁

2023-03-30 c4882b43fce3c32cb0ce3c9fc2c164f0ce0e8213

export

 funasr/export/export_model.py

@@ -205,6 +205,13 @@
            model, punc_train_args = PUNCTask.build_model_from_file(
                punc_train_config, punc_model_file, 'cpu'
            )
        elif mode.startswith('punc_VadRealtime'):
            from funasr.tasks.punctuation import PunctuationTask as PUNCTask
            punc_train_config = os.path.join(model_dir, 'config.yaml')
            punc_model_file = os.path.join(model_dir, 'punc.pb')
            model, punc_train_args = PUNCTask.build_model_from_file(
                punc_train_config, punc_model_file, 'cpu'
            )
        self._export(model, tag_name)
            


 funasr/export/models/encoder/sanm_encoder.py

@@ -9,6 +9,21 @@
from funasr.modules.positionwise_feed_forward import PositionwiseFeedForward
from funasr.export.models.modules.feedforward import PositionwiseFeedForward as PositionwiseFeedForward_export

def subsequent_mask(size, device="cpu", dtype=torch.bool):
    """Create mask for subsequent steps (size, size).

    :param int size: size of mask
    :param str device: "cpu" or "cuda" or torch.Tensor.device
    :param torch.dtype dtype: result dtype
    :rtype: torch.Tensor
    >>> subsequent_mask(3)
    [[1, 0, 0],
     [1, 1, 0],
     [1, 1, 1]]
    """
    ret = torch.ones(size, size, device=device, dtype=dtype)
    return torch.tril(ret, out=ret)

class SANMEncoder(nn.Module):
    def __init__(
        self,
@@ -150,10 +165,11 @@
    
    def prepare_mask(self, mask):
        mask_3d_btd = mask[:, :, None]
        sub_masks = subsequent_mask(mask.size(-1))
        if len(mask.shape) == 2:
            mask_4d_bhlt = 1 - mask[:, None, None, :]
            mask_4d_bhlt = 1 - sub_masks[:, None, None, :]
        elif len(mask.shape) == 3:
            mask_4d_bhlt = 1 - mask[:, None, :]
            mask_4d_bhlt = 1 - sub_masks[:, None, :]
        mask_4d_bhlt = mask_4d_bhlt * -10000.0
        
        return mask_3d_btd, mask_4d_bhlt
@@ -161,6 +177,7 @@
    def forward(self,
                speech: torch.Tensor,
                speech_lengths: torch.Tensor,
                vad_mask: torch.Tensor,
                ):
        speech = speech * self._output_size ** 0.5
        mask = self.make_pad_mask(speech_lengths)
@@ -173,8 +190,12 @@
        encoder_outs = self.model.encoders0(xs_pad, mask)
        xs_pad, masks = encoder_outs[0], encoder_outs[1]
        
        encoder_outs = self.model.encoders(xs_pad, mask)
        xs_pad, masks = encoder_outs[0], encoder_outs[1]
        # encoder_outs = self.model.encoders(xs_pad, mask)
        for layer_idx, encoder_layer in enumerate(self.model.encoders):
            if layer_idx == len(self.model.encoders) - 1:
                mask = (mask[0], vad_mask)
            encoder_outs = encoder_layer(xs_pad, mask)
            xs_pad, masks = encoder_outs[0], encoder_outs[1]
        
        xs_pad = self.model.after_norm(xs_pad)
        

 funasr/export/models/vad_realtime_transformer.py

@@ -21,7 +21,9 @@
        **kwargs,
    ):
        super().__init__()

        onnx = False
        if "onnx" in kwargs:
            onnx = kwargs["onnx"]

        self.embed = model.embed
        if isinstance(model.encoder, SANMVadEncoder):
@@ -53,12 +55,13 @@

    def get_dummy_inputs(self):
        length = 120
        text_indexes = torch.randint(0, self.embed.num_embeddings, (2, length))
        text_lengths = torch.tensor([length-20, length], dtype=torch.int32)
        return (text_indexes, text_lengths)
        text_indexes = torch.randint(0, self.embed.num_embeddings, (1, length))
        text_lengths = torch.tensor([length], dtype=torch.int32)
        vad_mask = torch.ones(length, length)[None, None, :, :]
        return (text_indexes, text_lengths, vad_mask)

    def get_input_names(self):
        return ['input', 'text_lengths']
        return ['input', 'text_lengths', 'vad_mask']

    def get_output_names(self):
        return ['logits']
@@ -66,14 +69,9 @@
    def get_dynamic_axes(self):
        return {
            'input': {
                0: 'batch_size',
                1: 'feats_length'
            },
            'text_lengths': {
                0: 'batch_size',
            },
            'logits': {
                0: 'batch_size',
                1: 'logits_length'
            },
        }